Merge pull request #27 from t81dev/phase-18-semantic-optimization-2052748221906594406

Phase 18 — Ternary Workload Maturation & Measurement Plane

Author: t81dev

.gitignore

@@ -1,5 +1,10 @@
 # --- Compiled Binaries & Objects ---
 bin/*
+benchmarks/synthetic/saturation
+benchmarks/synthetic/density
+benchmarks/gemm/gemm_bench
+benchmarks/lstm/lstm_bench
+benchmarks/temporal/lstm_persistent
 obj_dir/*
 *.o
 *.a

benchmarks/Makefile

@@ -2,15 +2,34 @@ CC=gcc
 CFLAGS=-O3 -I../include -I../src
 LDFLAGS=-L../bin -ltfmbs_device -lpthread
 
-TARGETS=layer1_synthetic layer2_kernels
+# Reorganized Phase 18 Structure
+SYNTHETIC_BINS=synthetic/saturation synthetic/density
+GEMM_BINS=gemm/gemm_bench
+LSTM_BINS=lstm/lstm_bench
+TEMPORAL_BINS=temporal/lstm_persistent
 
-all: $(TARGETS)
+ALL_BINS=$(SYNTHETIC_BINS) $(GEMM_BINS) $(LSTM_BINS) $(TEMPORAL_BINS)
 
-layer1_synthetic: layer1_synthetic.c
+all: $(ALL_BINS)
+
+synthetic/%: synthetic/%.c
+	$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)
+
+gemm/%: gemm/%.c
 	$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)
 
-layer2_kernels: layer2_kernels.c
+lstm/%: lstm/%.c
 	$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)
 
+temporal/%: temporal/%.c
+	$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)
+
+bench: all
+	@echo "Running Phase 18 Benchmarks..."
+	@for bin in $(ALL_BINS); do \
+		echo "Running $$bin..."; \
+		LD_LIBRARY_PATH=../bin ./$$bin; \
+	done
+
 clean:
-	rm -f $(TARGETS)
+	rm -f $(ALL_BINS)

benchmarks/density_metrics.csv

@@ -0,0 +1,3 @@
+zero_skips,total_ops,active_ops,mem_reads,mem_writes,broadcasts,residency_hits,residency_misses,cycles,fabric_cost,semantic_efficiency
+524288,524288,0,105061,2048,6,18,0,8738,541701.00,0.0000
+524288,524288,0,105061,2048,6,18,0,18,8738,541701.00,0.0000

benchmarks/gemm/gemm_bench.c

@@ -0,0 +1,36 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <time.h>
+#include "tfmbs_api.h"
+
+void benchmark_gemm_kernel() {
+    printf("\n--- [GEMM] Kernel Benchmark ---\n");
+    int rows = 1024, cols = 1024;
+
+    int8_t* host_w = malloc(rows * cols);
+    for(int j=0; j<rows*cols; j++) host_w[j] = (rand()%3)-1;
+
+    tfmbs_tensor_t w = tfmbs_tensor_bind(host_w, rows * cols, 1);
+    tfmbs_tensor_t i = tfmbs_tensor_bind(NULL, cols, 1);
+    tfmbs_tensor_t o = tfmbs_tensor_bind(NULL, rows * sizeof(int32_t), 0);
+
+    for(int n=0; n<5; n++) {
+        fabric_handle_t h = tfmbs_gemm(&w, &i, &o, rows, cols);
+        tfmbs_sync(h);
+
+        fabric_metrics_t m;
+        fabric_get_metrics(&m);
+        printf("Run %d: Cost: %.2f, Efficiency: %.4f, Residency Hits: %lu\n",
+               n, m.fabric_cost, m.semantic_efficiency, m.residency_hits);
+        tfmbs_dump_metrics_csv("gemm_metrics.csv");
+    }
+
+    tfmbs_tensor_release(&w); tfmbs_tensor_release(&i); tfmbs_tensor_release(&o);
+    free(host_w);
+}
+
+int main() {
+    benchmark_gemm_kernel();
+    return 0;
+}

benchmarks/gemm_metrics.csv

@@ -0,0 +1,3 @@
+zero_skips,total_ops,active_ops,mem_reads,mem_writes,broadcasts,residency_hits,residency_misses,cycles,fabric_cost,semantic_efficiency
+349272,1048576,699304,209919,4096,5,15,0,17476,1781677.00,0.3925
+349272,1048576,699304,209919,4096,5,15,0,15,17476,1781677.00,0.3925

benchmarks/layer1_synthetic.c

@@ -0,0 +1,39 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <time.h>
+#include "tfmbs_api.h"
+
+void benchmark_lstm_kernel() {
+    printf("\n--- [LSTM] Kernel Benchmark ---\n");
+    int h_size = 512;
+    int i_size = 512;
+    int rows = 4 * h_size;
+    int cols = i_size + h_size;
+
+    int8_t* host_w = malloc(rows * cols);
+    for(int j=0; j<rows*cols; j++) host_w[j] = (rand()%3)-1;
+
+    tfmbs_tensor_t w = tfmbs_tensor_bind(host_w, rows * cols, 1);
+    tfmbs_tensor_t i = tfmbs_tensor_bind(NULL, cols, 1);
+    tfmbs_tensor_t o = tfmbs_tensor_bind(NULL, rows * sizeof(int32_t), 0);
+
+    for(int n=0; n<5; n++) {
+        fabric_handle_t h = tfmbs_lstm_step(&w, &i, &o, h_size, i_size);
+        tfmbs_sync(h);
+
+        fabric_metrics_t m;
+        fabric_get_metrics(&m);
+        printf("Run %d: Cost: %.2f, Efficiency: %.4f, Residency Hits: %lu\n",
+               n, m.fabric_cost, m.semantic_efficiency, m.residency_hits);
+    }
+
+    tfmbs_tensor_release(&w); tfmbs_tensor_release(&i); tfmbs_tensor_release(&o);
+    free(host_w);
+}
+
+int main() {
+    benchmark_lstm_kernel();
+    tfmbs_dump_metrics_csv("lstm_metrics.csv");
+    return 0;
+}

benchmarks/layer2_kernels.c

@@ -0,0 +1,3 @@
+zero_skips,total_ops,active_ops,mem_reads,mem_writes,broadcasts,residency_hits,residency_misses,cycles,fabric_cost,semantic_efficiency
+699703,2097152,1397449,419634,8192,0,15,0,3072,3561155.00,0.3924
+699703,2097152,1397449,419634,8192,0,15,0,0,3072,3561155.00,0.3924

benchmarks/lstm/lstm_bench.c

@@ -0,0 +1,3 @@
+zero_skips,total_ops,active_ops,mem_reads,mem_writes,broadcasts,residency_hits,residency_misses,cycles,fabric_cost,semantic_efficiency
+0,1048576,1048576,209919,4096,3,12,0,17476,2130945.00,0.4921
+0,1048576,1048576,209919,4096,3,12,0,6,17476,2130945.00,0.4921